Using Genomics to Explore if Hot Flushes Cause Insomnia

If you’re here for the answer, it's at the bottom of this page. However, if you have any interest in biology/genomics, I believe the method of arriving at that answer is just as, if not more interesting than the answer itself. Here’s how I (with the Reproductive Genomics Team) proved that hot flushes cause insomnia¹, purely from genomic data.

^{1 Strong evidence for at least partial causality - you can't quite be 100% sure in science, can you?}

Menopause is the permanent cessation of menstruation, a phase in a typical biological woman’s life². It features distinct changes in hormone levels such as those of oestrogen, oestradiol, and progesterone, and, as a result or otherwise, is accompanied by a variety of symptoms that can be disruptive to daily life.

Insomnia, on the other hand, is a lot easier to understand. It is simply the difficulty of falling asleep/maintaining sleep, despite having the opportunity to do so.

Hot flushes are a highly disruptive symptom of menopause in up to 80% of women, and can last well past menopause. Essentially, a misregulation of internal body temperature leads to vasodilation, releasing body heat, thus leading to the individual suddenly feeling extremely, sometimes even unbearably hot.

^{2 Defined here as having 2 'X' chromosomes and no 'Y' chromosomes. It's not a perfect definition, but I hope it gets the point across!}

Insomnia is another key symptom of menopause, and often occurs along with hot flushes. It has long been thought that menopausal insomnia might, at least in part, be caused by hot flushes. Here’s the core idea:

→  Hot flushes that trigger during an individual’s sleep disrupt the sleep cycle
→  This disruption reduces sleep quality
→  The individual may also be woken up by the hot flush

This is called the 'Domino Theory' - one thing causes another. Sounds straightforward, right? But it gets tricky real fast.

Right now, this is just that: a theory. So far, all we really have is an association, and associations can be misleading. For example, it could be that psychosocial stress (during menopause) causes both insomnia and hot flushes independently. As a result, insomnia and hot flushes appear together, but they don't cause one another. This might seem like splitting hairs, but it's important.

For example, future research focusing on insomnia might target hot flushes with a drug, hoping to alleviate insomnia. But if insomnia is not caused by hot flushes, this would not have the intended effect.

What we need is causal proof. And the gold standard for causal proofs is the randomised control experiment.

The randomised control experiment is as simple as it is clever, and is most famously employed when testing if a drug is responsible for an effect, or is just associated with the effect. Here's how it goes:
→ Randomly allocate people into 2 groups
→ Administer one group (called the focus group) with the drug you're testing
→ Administer another group (called the control group) with a placebo pill (a pill that doesn't do anything)
→ Check to see if the focus group behaves differently from the control. Since the only difference between them was the drug itself, the difference must be because of the drug.

For example, if both groups showed the same effects, then a possible explanation is that the drug itself doesn't have any effect. The drug is a dud. But if there is a difference, e.g. side effects or the intended effect, this means that the drug has had an effect.

Essentially, we can turn a naïve association into a causal association, i.e. where we can say with some confidence that one thing causes another.

But hold on! While this works well for testing drugs, we can't exactly do this with hot flushes and insomnia. For one, we can't 'administer' hot flushes to a group and hold another one as the 'control', to see which group develops insomnia. Even if we found a way to trigger hot flushes artificially (e.g. by changing the hormone levels), there are obvious ethical considerations.

So... that's it. The project can't be done, right? Not quite so fast.

Mendelian Randomisation is a genomics technique that uses genetic inheritance as a population-wide randomised control experiment. If that sounds complicated, don't bother rereading that sentence. The next few paragraphs are all dedicated to explaining this fantastic technique - honestly, I've wondered more than once how someone came up with this.

Here's the gist:
→ Imagine there's a gene 'A' that you know is responsible for hot flushes
→ At gamete formation, the genes passed to offspring are random. As such, some people inherit gene A, while others don't
→ Gather a bunch of people
→ Check how many have Gene A - this is the 'focus' group
→ Check how many don't have Gene A - this is the 'control' group

Now, here's the clever bit: we check for an association between Gene A and Insomnia.

Because Gene A causes Hot Flushes, we're gonna have it be our proxy. Additionally, because we're only working with the genetics, which don't change over a person's life, we don't have to worry about factors such as stress muddying up our results.

Because Gene A was divided into the population randomly, and we're using it as a proxy for hot flushes, we can say the hot flushes were 'distributed' randomly. Now, we're essentially just doing a population-wide randomised control experiment!

But wait! What we've really found at this point is that Gene A causes insomnia. How do we know it's because of hot flushes? Fantastic question! The stats and maths can only go so far, so it's time for a bit of biology.

The Mendelian Randomisation test assumes that there is no pleiotropy. This means that, when using this test, we need to ensure that we pick a 'Gene A' that we are sure doesn't have any pleiotropic effects on insomnia. Luckily for me, this is where the lab had already done the groundwork. Enter TACR3, a gene that codes for the tachykinin receptor, and has a mountain of research behind it that points towards it causing hot flushes. Additionally, there has been no evidence to suggest that it is involved in pathways directly leading to insomnia. This makes it a fantastic candidate.

But that's not enough. So far, I've been saying 'Gene A', but in the real world, we use a set of genes, that is to say, 'Gene A', 'Gene B', 'Gene C', and so on, and test for an association with all of them, together. My first couple of weeks of this project were spent looking into the biology of the most significant of these genes, looking to see if there were any effects we needed to be aware of. We eliminated a few genes for various reasons I won't get into now, but the TLDR is that we got a set of genes that we believe satisfy these conditions.

I say believe, and I think it's key that that's remembered. Often, when the big headlines come out, it's that 'A causes B', or 'Hey, watch out if you've experienced this!'. When it really comes down to it, there are a lot of judgment calls researchers make that can shape the results - I've realised this doubly while working with the team here. Often, when I was excited to plough ahead and feed the numbers into the magic stats tests, the team had me rein in my horses and explore data to make sure that what we were finding wasn't abstract mathematical associations, but real biology. I feel like what's presented in the final line of the results, in science at least, always requires a grain of salt, because behind all those numbers and confidence scores are human decisions made to the best of their abilities.

I said 'gather a bunch of people', but we can't exactly just 'gather' a bunch of people. For these sorts of association tests, the combined groups were on the order of over 100,000 people. Yep, not hundreds, not thousands, not even tens of thousands. The scale is orders of magnitude higher than some of the leading research, such as the SWAN experiment that came previously.

So, how exactly do we achieve this? Luckily, we didn't have to collect the data ourselves. The UKBioBank is a national database of real people's genomes and health records across various points in their lives. From this database, we managed to get the data we needed quite easily - that's what I'd say, but really, this was the step that took the majority of my internship.

See, insomnia is a fuzzy term. How do you differentiate it from sleeplessness? It's not a simple yes or no, either; some people can be more insomniac than others.

Don't worry - I'll spare you the details, though if you're interested, you can read the full report here. The take-home is that, through a lot of meetings, discussions, and fiddling around with Python code, we distilled a group of individuals we strongly believe are either definitely insomniatic or definitely not insomniatic (the latter for the controls).

The final step! Plugging in all the data into the R script (the code performed the test!), and waiting for several intense minutes, my laptop breathed a proud sigh, brought out some vintage wine, raised its glass and announced long-awaited output.

No, that's not what really happened, in true science fashion, the laptop just outputted the result as a table without any fanfare. Was I expecting some dramatic reveal? Not sure why, but absolutely! I should really edit the code to say 'Congratulations! Here's the result!' or something. Still, the contents of that table were exactly what I was looking for!

With a p-value (the lower the value, the more confident we are of the result) on the order of 1x10^-5 (really anything below 0.05 for this test was enough), we found a genetic, causal correlation, free of any meddling lifestyle factors, age, and stress. In fact, we found that hot flushes double your risk of insomnia³.

^{3The analysis ignored the direction of effect when creating the polygenic risk score. This is instead the most likely correct interpretation of the results. Strictly speaking, we know that having hot flushes changes your chances of having Insomnia by 2 times - not necessarily an increase though, as this was not seperately tested for.}

This shouldn't be taken to mean that hot flushes are the only cause for insomnia. Quite obviously, factors such as age and stress do play a role.

This also does not mean that insomnia does not cause hot flushes. This is very much possible, though I can't comment on it since this test didn't explore that.

It also does not prove the Domino Theory - though, without other explanations, it does support the theory.

Hot flushes appear to increase the risk of insomnia.

Yep, for all that work, I still have to include the word 'appear'.

Science is cautious, and despite the amazing p-values and ingenious statistical tests, the results do reflect this caution. This project isn't going to change how we treat insomnia or hot flushes right off the bat. That can sound like a negative, but really, if anything, I feel even more confident in the science that does reach me today.

Each medicine that you take, each policy that is made is the result of work under such fantastic scrutiny. As an undergraduate student, research is often presented to me as this slow, tedious process, but really, it's dynamicity with guardrails. Exploration, but with caution.

To me, this internship's take-home message wasn't just the research output, or the statistical tests I learnt to work with, but a more holistic idea of the work and creativity that goes into those numbers we can read in the blink of an eye.

As for future work, confirming the direction of effect (positive, negative) would be a good start, and from there expanding to 2 sample MRs and testing for bidirectional effects. As for me, I'm going to continue exploring the various topics I came across over the past 2 months, and keep building more tools to explore them. Hopefully, I'll have something to share here, again :)